Re: Lockdep complains about commit 1331e7a1bb ("rcu: Remove_rcu_barrier() dependency on __stop_machine()")

From: Jiri Kosina
Date: Wed Oct 03 2012 - 04:24:58 EST


On Wed, 3 Oct 2012, Srivatsa S. Bhat wrote:

> On 10/03/2012 01:13 PM, Jiri Kosina wrote:
> > On Wed, 3 Oct 2012, Srivatsa S. Bhat wrote:
> >
> >>>>> CPU 0 CPU 1
> >>>>> kmem_cache_destroy()
> >>>>
> >>>> What about the get_online_cpus() right here at CPU0 before
> >>>> calling mutex_lock(slab_mutex)? How can the cpu_up() proceed
> >>>> on CPU1?? I still don't get it... :(
> >>>>
> >>>> (kmem_cache_destroy() uses get/put_online_cpus() around acquiring
> >>>> and releasing slab_mutex).
> >>>
> >>> The problem is that there is a CPU-hotplug notifier for slab, which
> >>> establishes hotplug->slab.
> >>
> >> Agreed.
> >>
> >>> Then having kmem_cache_destroy() call
> >>> rcu_barrier() under the lock
> >>
> >> Ah, that's where I disagree. kmem_cache_destroy() *cannot* proceed at
> >> this point in time, because it has invoked get_online_cpus()! It simply
> >> cannot be running past that point in the presence of a running hotplug
> >> notifier! So, kmem_cache_destroy() should have been sleeping on the
> >> hotplug lock, waiting for the notifier to release it, no?
> >
> > Please look carefully at the scenario again. kmem_cache_destroy() calls
> > get_online_cpus() before the hotplug notifier even starts. Hence it has no
> > reason to block there (noone is holding hotplug lock).
> >
>
> Agreed.
>
> > *Then* hotplug notifier fires up, succeeds obtaining hotplug lock,
>
> Ah, that's the problem! The hotplug reader-writer synchronization is not just
> via a simple mutex. Its a refcount underneath. If kmem_cache_destroy() incremented
> the refcount, the hotplug-writer (cpu_up) will release the hotplug lock immediately
> and try again. IOW, a hotplug-reader (kmem_cache_destroy()) and a hotplug-writer
> (cpu_up) can *NEVER* run concurrently. If they do, we are totally screwed!
>
>
> Take a look at the hotplug lock acquire function at the writer side:
>
> static void cpu_hotplug_begin(void)
> {
> cpu_hotplug.active_writer = current;
>
> for (;;) {
> mutex_lock(&cpu_hotplug.lock);
> if (likely(!cpu_hotplug.refcount)) <================ This one!
> break;
> __set_current_state(TASK_UNINTERRUPTIBLE);
> mutex_unlock(&cpu_hotplug.lock);
> schedule();
> }
> }

I acutally just came to the same conclusion (7 hours of sleep later, the
mind indeed seems to be brighter ... what a poet I am).

Lockdep doesn't know about this semantics of cpu_hotplug_begin(), and
therefore gets confused by the fact that mutual exclusion is actually
achieved through the refcount instead of mutex (and the same apparently
happened to me).

So right, now I agree that the deadlock scenario I have come up with is
indeed bogus (*), and we just have to annotate this fact to lockdep
somehow.

And I actually believe that moving the slab_mutex around in
kmem_cache_destroy() is a good anotation (maybe worth a separate comment
in the code), as it not only makes the lockdep false positive go away, but
it also reduces the mutex hold time.

(*) I have seen machine locking hard reproducibly, but that was only with
additional Paul's patch, so I guess the lock order there actually was
wrong

Thanks!

--
Jiri Kosina
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/